Unsupervised learning to identify symptom clusters in older adults undergoing chemotherapy.

introductionUnsupervised machine learning (ML) approaches such as clustering have not been commonly applied to patient-reported data. This study describes ML methods to explore and describe patient-reported symptom trajectories in older adults receiving chemotherapy. MATERIALS AND

methodsThis secondary analysis of prospectively collected data from the GAP 70+ Trial (NCT02054741; PI: Mohile) collected patient-reported symptoms at baseline (pre-chemotherapy), six weeks, three months, and six months. Complete patient-reported symptom data were available for at least one timepoint for 708/718 patients (98.6 %). Correlation analysis was performed on all symptom items. Multiple clustering algorithms were applied to selected baseline symptoms as an exploratory analysis, using gap statistic and elbow plots to understand optimal cluster numbers for each algorithm. Silhouette scores and t-stochastic neighbor embedding (t-SNE) plots were generated for each algorithm. Hierarchical agglomerative clustering was applied to symptoms at each timepoint, and clusters generated for each timepoint were examined longitudinally utilizing statistical measures, violin plots, and a Sankey diagram.

resultsTwenty-six patient-reported items were used for clustering analyses, representing symptom severity and interference. There was significant variability in how different unsupervised learning algorithms clustered the baseline symptom data. Silhouette scores ranged from -0.22 (OPTICS) to 0.16 (BIRCH). Examining agglomerative clustering across timepoints, cluster composition was largely driven by the symptom sum score (i.e., adding the Likert-scale scores). Most patients had "low" symptoms at baseline that remained low, but symptom trajectory was otherwise heterogeneous. A small number of patients had high hand-foot/neuropathy symptoms (but low other symptoms) at six weeks, and another small cluster had high mucosal toxicity at six months. Despite specific symptom patterns in these small clusters, chemotherapy regimens varied. DISCUSSION: Unsupervised machine learning techniques may be helpful to understand longitudinal patient-reported data such as symptoms. They permit data-driven exploration, which may uncover patterns to inform hypotheses or further analysis (e.g., outcome prediction). Results of clustering analyses should be validated through further hypothesis-driven analysis. In this analysis, it was challenging to uncover consistent symptom patterns, though it suggests symptom composite (sum) scores may warrant further investigation. Clinicians should understand the philosophy, strengths, and limitations of an unsupervised machine learning approach applied to patient data.